Table 1.

Listing of features within feature groups.

Survey categories	Mobility features	Sociability features	Cognition features	Screen time features	Sleep features
1. Anxiety	1. Time spent at home	1. Number of outgoing texts	1. Jewels Beta A	1. Screen time	1. Sleep duration
2. Social	2. Distance traveled	2. Total outgoing text length	2. Jewels Beta B	2. Session time
3. Medication	3. Radius of gyration	3. Texting out-degree		3. Checks
4. Psychosis	4. Maximum diameter	4. Number of incoming texts
5. Depression	5. Maximum distance from home	5. Total incoming text length
6. Sleep	6. Number of significant locations	6. Texting in-degree
	7. Average flight length	7. Texting reciprocity
	8. Standard deviation of flight length	8. Texting responsiveness
	9. Average flight duration	9. Number of outgoing calls
	10. Standard deviation of flight duration	10. Total outgoing call duration
	11. Fraction of the day spent stationary	11. Call out-degree
	12. Significant location entropy	12. Number of incoming calls
	13. Minutes of GPS data missing	13. Total incoming call durations
	14. Physical circadian rhythm	14. Call in-degree
	15. Physical circadian rhythm stratified	15. Call reciprocity
		16. Call responsiveness

Features are calculated for each participant for each day data is available. Cognition features are derived from a modified Cox proportional hazard model previously reported on¹³. Screen time features are derived from power state sensor data and checking behavior specifically (phone sessions lasting < 15 s) has been shown to be a proxy for problematic smartphone usage²⁰. Sleep duration estimates were generated from sensor fusion of GPS, accelerometer, and screen usage events, by defining a significance threshold for filtration of extraneous observations. For GPS, the Euclidian distance between two events was taken; for accelerometer, the difference in magnitude between two events was taken; for screen usage, the duration of time the screen was turned on or the device was unlocked was taken. These three streams were binned in 24-h intervals starting at 12:00 (noon), and for all three sensors, the median of both sleep time and wake time were taken per bin to create a single sleep period.